-
https://git.ecker.tech/ aims to provide a place to share my efforts while maintaining true ownership of my code, as I do not trust GitHub.
XMR: 4B9TQdkAkBFYrbj5ztvTx89e5LpucPeTSPzemCihdDi9EBnx7btn8RDNZTBz2zihWsjMnDkzn5As1LU6gLv3KQy8BLsZ8SG
- Joined on
2022-10-10
The Sample slider in the Generate tab will only affect how many (autoregressive) samples you want.
The progress bar on the right will report [Current Batch / Total Batches], where `Total…
Oops, a slight oversight. Should be fixed in commit fefc7aba0388d061a2db37cd74ebef7a9172439b.
Apologies for taking so long to get around to this. A mix of everything interrupting me, and trying to follow through with a possible breakthrough with training on lower VRAM cards.
Added…
Should be fixed in commit 9e64dad785bc6d6e0e948784b27028d5d4746e25. I can't do any testes for it on my 2060 since I can't increase it's batch size all that much for higher numbers.
9e64dad785
clamp batch size to sample count when generating for the sickos that want that, added setting to remove non-final output after a generation, something else I forgot already
Setting samples in Gradio to any number divides the output by 12, 16 etc?
That's from it automatically deducing your batch size. On DirectML, there doesn't seem to be an (easy) way to get…
Glad to hear.
holy moly is ROCm way faster
Yeah, DirectML unfortunately is still in its infancy, and I need to do some patches where it'll do some calculations on CPU due to some…
f119993fb5
explicitly use python3 because some OSs will not have python alias to python3, allow batch size 1
Right, some installs will not have python as an alias. I'll push a fix out, although I swore I did this already before, but it might have gotten reverted during migration from the old repo.
As I mentioned in mrq/ai-voice-cloning#17 (comment), I added an experimental way to train fully at half-precision that will convert the original model to one…
I've added/exposed a very experimental training setting: Half Precision in commit 8a1a48f31e30957196213a193c9aab45f1c25520. It'll convert the original training model to float16 and enable…
I tried again on an RTX4000 (8GiB VRAM) with:
- half-precision training
- experimental half-precision training model
- batch size 1 (which surprisingly works)
- mega batch factor 1 (changing…
8a1a48f31e
Added very experimental float16 training for cards with not enough VRAM (10GiB and below, maybe) \!NOTE\! this is VERY EXPERIMETNAL, I have zero free time to validate it right now, I'll do it later